V4-Flash by default for cheap iteration; /pro lifts a single turn to V4-Pro

Node ≥ 22 on macOS / Linux / Windows

In long sessions the bill typically lands at ~1/3 of comparable generic tooling.

$0.07 /Mtok in · $0.014 /Mtok cached

long sessions hold 90%+ cache hit and input-token cost collapses to ~1/5

Perceptual BD-rates are based on human ratings from a large-scale subjective study

search over millions of model configurations to jointly optimize over perceptual quality and on-device runtime

Based on large-scale subjective user studies

faster than most top ML-based codecs run on a V100 GPU

on an iPhone 17 Pro Max, it encodes 12MP images as fast as 230ms, and decodes them in 150ms

existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions.

error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes.

agents succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django).

Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, while some weaker configurations approach zero.

Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline.

However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings.

the model alone is no longer the product

if you can effectively posttrain a model to only meaningfully perform with your closed source agent, then you get to funnel the majority of users to your agent at the expense of your model/API co-opetition

The quote is a big reversal of stance from a position ~uniformly held by anyone who worked at Team Big Model, including his previous head of OpenAI Labs

agentic systems can be designed to call on such tools when they might be useful

For the next decade or so, we should think about AI as this amazing tool to help scientists

general-purpose reasoning model in the vein of GPT-5.5

Google fellow John Jumper, who won the Nobel for AlphaFold, is now working on AI coding, not on science-specific AI tools

the best data filter may be **no filter**, with projections suggesting the crossover for internet-scale pools lands around **1e30 FLOPs**

Hark raised $700M

Modal raised big

turbopuffer crossed $100M run-rate

Claude Opus 4.7 has been used to patch over 2,100 vulnerabilities

on average, a high- or critical-severity bug found by Mythos Preview takes two weeks to patch

90.6% (1,587) have proved to be valid true positives, and 62.4% (1,094) were confirmed as either high- or critical-severity

Mythos Preview has found what it estimates are 6,202 high- or critical-severity vulnerabilities in these projects (out of 23,019 in total)

their rate of bug-finding has increased by more than a factor of ten

we and our approximately 50 partners have used Claude Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities

We have been watching what developers have built on Claude over the last few years, which made bringing our teams together an easy decision.

Anthropic created MCP to make agent connectivity possible.

Agents are only as useful as what they can connect to.

SDKs deserve as much care as the APIs they wrap.

The frontier of AI is shifting from models that answer to agents that act—and agents are only as capable as the systems they can reach.

In my opinion this paper demonstrates that current AI models go beyond just helpers to human mathematicians – they are capable of having original ingenious ideas, and then carrying them out to fruition.

The key ingredients of the construction come from a very different part of mathematics known as algebraic number theory, which studies concepts like factorization in extensions of the integers known as algebraic number fields.

The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular.

An internal OpenAI model has disproved this longstanding conjecture, providing an infinite family of examples that yield a polynomial improvement.

The result is also notable for how it was found. The proof came from a new general-purpose reasoning model... In this case, it produced a proof resolving the open problem.

The precise argument uses tools such as infinite class field towers and Golod–Shafarevich theory to show the number fields required for the argument actually exist. These ideas were well-known to algebraic number theorists, but it came as a great surprise that these concepts have implications for geometric questions in the Euclidean plane.

The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular.

6.4 billion from operations on just 3.2 billion

orbital AI compute satellites as early as 2028

multiple trillions of parameters

$18.5 billion in purchases

$43 billion in privately held stakes

first operating profit

Our National Partnerships for AI Working with governments worldwide to benefit people through frontier AI

Veo Generate cinematic video with audio

AlphaEvolve Design advanced algorithms for math and applications in computing

SIMA 2 An agent that plays, reasons, and learns with you in virtual 3d worlds

Gemini Robotics Perceive, reason, use tools and interact

Gemini Omni Create anything from anything

AlphaEvolve Design advanced algorithms for math and applications in computing

Gemini Robotics Perceive, reason, use tools and interact

Lyria Generate high fidelity music and audio

Imagen Generate high-quality images from text

Veo Generate cinematic video with audio

Gemini Build intelligent agents

Gemini Omni Create anything from anything

Anthropic leads OpenAI in business adoption, according to Ramp.

annualized revenues approaching $50 billion – a fivefold increase in as many months.

90% of finance reporting is now AI-driven as well.

Chinese AI labs have developed an efficiency moat that may define the AI market's development over the coming years.

there are around 10,000 people— founders and employees at companies like OpenAI, Anthropic, and Nvidia — that have 'hit retirement wealth of well above $20M'

many software engineers feel that their life's skill is no longer useful

the same technology is both the lottery ticket & the thing eating your fallback

the divide in outcomes is the worst I've ever seen

The vibes around the current AI boom aren't great, even in the tech industry

Another secondary summary gives Humanity’s Last Exam: 64.7% vs 53.1%, possibly under different setup/effort/tool conditions.

Anthropic explicitly says Mythos Preview is available to launch partners in Project Glasswing, not general users... This triggered discussion of “API hoarding” and a new closed-access elite tier.

The interpretation that Anthropic has “the mandate” or is undervalued at $380B is an investor thesis, not a confirmed market fact.

A key subtext in the tweets is that high-margin enterprise/coding/cyber workloads may now be sufficient to support frontier labs without broad public access to their best models. This becomes more plausible if Anthropic’s revenue is indeed compounding as fast as posters claim.

We’ve done a focused news summary run below, for those who desire more detail.

If a master tactician wanted to further competitive narratives vs a potential IPO, you would be hard pressed to find a better idea than Claude Mythos... and now formally confirmed to be too dangerous to release GA, instead only restricted to 40 partners under an urgent new “Project GlassWing”

Against the backdrop of OpenAI announcing $24B ARR, stalled ChatGPT growth and coincidental personnel moves in CEO, COO, and CMO and sensationalist rumors with CFO, this week’s events in Anthropic announcing a massive jump from $19B ARR in March to $30B ARR in April seems like a VERY strategic jab, especially considering known differences in revenue recognition, but the differential rate of growth and higher cost efficiency is undeniable… only for today to step it up a notch.

A photo of a scribbled note becomes an interactive to-do list; a paused frame in a travel video becomes a booking link for that cool-looking restaurant.

In everyday interactions with each other, humans rarely speak in long, detailed paragraphs. We might say, "Fix this", "Move that here", or "What does this mean?" — while relying on physical gestures and our shared context to fill in any gaps

For decades, computers have only tracked where we are pointing. AI can now also understand what the user is pointing at. This transforms pixels into structured entities, such as places, dates, and objects

the pointer has barely evolved in more than half a century.

because a typical AI tool lives in its own window, users need to drag their world into it. We want the opposite: intuitive AI that meets users across all the tools they use, without interrupting their flow.

Shaping the future of AI interaction by reimagining the mouse pointer — Google DeepMind

Domain-specific ECI scores can be used to compare performance relative to other model releases, but not to track the absolute performance or progress trends in different domains.

The SWE overperformance has been consistent across most generations, and remains in recent models.

The most extreme ratio observed is 4 math benchmarks to 2 SWE benchmarks.

All models included in our analysis have at least two scores in each domain, with an average of 3.2 SWE benchmark results and 3.4 math benchmark results.

Opus 4.6 and 4.7 both have Math-ECIs within 1 point of their general ECI, compared to larger gaps for earlier models.

On average Claude models have an SWE-ECI 2.7 points higher than their general ECI, and a Math-ECI 1.8 points lower.

We believe AI can meaningfully expand what's possible for the smallest businesses, including solo entrepreneurs.

Small businesses account for 44% of U.S. GDP and employ nearly half the private-sector workforce, but their adoption of AI has lagged behind larger enterprises.

Small businesses need AI that moves at the speed they do. With Canva powering content creation in Claude for Small Business, a business owner can go from idea to published, on-brand design in one flow

What we used to think were the constraints are just not constraints anymore. It's empowering. Hours of looking at stuff that doesn't matter are gone.

We don't train on your data by default on our Team and Enterprise Plans.

AI is the first technology that can finally close that gap, which is why we're launching Claude for Small Business

We intend to publish our thinking and decision-making as we do

The first of these will be released publicly later this year

In sub-Saharan Africa and India, we are creating AI-powered apps that support foundational literacy and numeracy programs

PwC will roll out Claude Code and Cowork starting with U.S. teams and expanding toward a global workforce of hundreds of thousands of professionals, establish a joint Center of Excellence, and train and certify 30,000 PwC professionals on Claude

KPMG and Anthropic announce a global alliance, with Claude integrated into KPMG's Digital Gateway platform and available to all 276,000+ employees

the nearly two billion people whose incomes depend on smallholder farming

commit $200 million in grant funding, Claude usage credits, and technical support for programs in global health, life sciences, education, and economic mobility over the next four years

building toward full-scale deployment across its 167,000-person workforce

the $100 million investment we made this year to back the services firms helping enterprises actually deploy AI

more than 5,000 leaders saw the alliance up close, with hands-on training enabling a wave of early adopters

Security work that took hours now takes minutes

Insurance underwriting that took 10 weeks now takes 10 days

cutting delivery times by up to 70%

a program to train and certify 30,000 PwC professionals on Claude

PwC will roll out Claude Code and Cowork starting with U.S. teams and expanding toward a global workforce of hundreds of thousands of professionals

a drag that is estimated to be more than $2 trillion

It's very enticing to say we're just going to replace everything with a chatbot, but it's not changing the bottom line.

Frankly, no customer ever just wants to talk to your chatbot.

Willis said there's no magic for innovating. Companies need to do the hard work of understanding how AI may or may not be useful for the desired outcome.

The deeper problem, he said, is that companies are treating AI itself as a solution rather than as a tool to help power the solution.

What company leaders face, he said, is not an innovation problem but an impatience problem.

the continued flood of AI reports has basically made the security list almost entirely unmanageable

Torvalds' remarks contrast with recent comments from fellow kernel maintainer Greg Kroah-Hartman, who recently told The Register that AI has become an increasingly useful tool for the FOSS community.

If you found a bug using AI tools, the chances are somebody else found it too.

AI tools are great, but only if they actually help, rather than cause unnecessary pain and pointless make-believe work.

AI detected bugs are pretty much by definition not secret, and treating them on some private list is a waste of time for everybody involved

People spend all their time just forwarding things to the right people or saying 'that was already fixed a week/month ago' and pointing to the public discussion.

the continued flood of AI reports has basically made the security list almost entirely unmanageable, with enormous duplication due to different people finding the same things with the same tools.

pluralism is most decisively made or unmade at the deployment-governance layer: interfaces, preference-data pipelines, and audit infrastructure.

We formalise a metric, the Pluralistic Repair Score (PRS), distinguishing principled revision from capitulation

the failure mode of contemporary RLHF-trained assistants is not insufficient coverage but sycophantic consensus

the collapse of disagreement at the interaction layer is not a narrow technical concern but a structural failure with distributive consequences.

We argue that aggregation alone is an incomplete primitive for deployed pluralistic alignment.

No IAM framework governs human privilege escalation and agent privilege escalation with the same rigor.

Most scanners track every CVE but cannot alert when a branch name exfiltrates a GitHub token through a container that developers trust by default.

Agents just made the cost of not doing it catastrophic.

It uses far more permissions than it should have, more than a human would, because of the speed of scale and intent.

The agent itself is the attack surface.

Static pattern matching loses to embedded prompts in legitimate review and Codespaces flows.

Threat actors are reverse engineering patches within 72 hours. If a customer doesn't patch within 72 hours of release, they're open to exploit.

Every attacker went for the credential, not the model.

AlphaEvolve has been used as a regular tool to optimize the design of the next generation of TPUs. It also helped discover more efficient cache replacement policies, achieving in two days what previously required a concerted, human-intensive effort spanning months.

AlphaEvolve began optimizing the lowest levels of hardware powering our AI stacks. It proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of our next-generation TPUs.

This optimization reduced 'write amplification'—the ratio of data written to storage versus the original request—by 20%. It also provided insights for new compiler optimization strategies that reduced the storage footprint of software by nearly 9%.

achieving 10% accuracy gains over their competitive manual model optimizations

doubling its training speed whilst improving model quality

reduced 'write amplification'—the ratio of data written to storage versus the original request—by 20%

finding 10.4% improvement in routing efficiency over the previous heavily optimized solutions — saving over 15,000 kilometers of distance travelled annually.

suggesting quantum circuits with 10x lower error than previous conventionally optimized baselines

the overall accuracy of predicting the risk of natural disaster—aggregated across 20 categories such as wildfires, floods, and tornadoes—was increased by 5%.

increase the ability of our trained Graph Neural Network (GNN) model to find feasible solutions for the problem from 14% to over 88%

achieving a 30% reduction in variant detection errors.

YouTube commenters started naming the robots Bob, Frank, and Gary yesterday, so we added name tags to each robot

If a robot has a software or hardware issue, it autonomously leaves for maintenance and another robot takes over.

If the robot gets stuck or the AI policy goes out of distribution, Helix triggers an automatic reset.

There is no teleoperation - every action comes directly from Helix-02

The robots are reasoning directly from camera pixels

Humans average around 3 seconds per package. F.03 is now around human parity.

When you stop using the agent, all the productivity benefit goes away... but the added maintenance costs don't!

If you want a productive team, you have to focus on their maintenance costs.

For every month you spend writing code, you'll spend some amount of time in the following year maintaining that code, and some in each year after that, forever, as long as that code exists.

occasionally even identifying the benchmark

Models sometimes recognize they're being evaluated

New research from @AISecurityInst and Goodfire

meaning safety benchmarks may not reflect real-world behavior

We show this verbalized eval awareness inflates safety scores

Models sometimes recognize they're being evaluated, occasionally even identifying the benchmark.

It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.

The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline

The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens

achieving gold-medal-level performance on mathematical and physics competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025.

achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025

Next-generation AI chips, such as Nvidia's Rubin, will shift to the 3nm process

of the roughly $30 billion year-over-year increase, around $20 billion came from HBM alone.

Total spending on components across the top four designers more than doubled from 2024 to 2025, rising from $22 billion to $52 billion.

The four designers consumed only ~11% of global leading-edge logic wafer capacity in 2024 and 2025.

The four designers still take roughly 80–85% of total CoWoS supply.

The top four designers collectively consumed nearly all of TSMC's CoWoS wafer output, leaving little headroom for other customers.

AI doesn't own state transitions. The Bubble Tea architecture has a beautiful idea: Update() is the only place state mutates, driven by messages.

AI generates this pattern because it's the shortest path from 'fetch data' to 'render table.'

The complexity was accumulating invisibly while the velocity metric said 'you're shipping!'

AI writes features, not architecture. The longer you let it drive without constraints, the worse the wreckage gets.

The velocity makes you think you're winning right up until the moment everything collapses simultaneously.

The tl;dr of this dev log is that I still need to be in the loop to make anything meaningful.

HTML can allow you to interact with the document, for example you might want to ask it to add sliders or knobs to adjust a design or allow you to tweak different options in the algorithm to see what happens. You can also ask it to let you copy these changes into a prompt to paste back into Claude Code.

The chance of someone actually reading your spec, report or PR writeup is much much higher if it's in HTML. HTML documents are much easier to read, Claude can organize the structure visually to be ideal to navigate with tabs, illustrations, links, etc. It can even be mobile responsive so you can read it differently based on your form factor.

HTML can convey much richer information compared to markdown. It can of course do simple document structure like headers and formatting, but it can also represent all sorts of other information such as: Tabular data using tables, Design data with CSS, Illustrations with SVG, Code snippets with script tags, Interactions using HTML elements with javascript + CSS, Workflows using SVG and HTML, Spatial data using absolute positions and canvases, Images using image tags

I find it difficult to read a markdown file of more than a hundred lines. I want richer visualizations, color and diagrams and I want to be able to share them easily.

Markdown has become the dominant file format used by agents to communicate with us. It's simple, portable, has some rich text capability and is easy for you to edit. Claude has even gotten surprisingly good at using ASCII to make diagrams inside of markdown files.

We recommend readers try out the interactive NLA demo hosted on Neuronpedia at this link. We have also released our code for other researchers to build on.

In more detail, suppose we have a language model whose activations we want to understand. NLAs work as follows. We make three copies of this language model: The target model is a frozen copy of the original language model that we extract activations from.

An auditor equipped with NLAs successfully uncovered the target model's hidden motivation between 12% and 15% of the time, even without access to the training data that implanted it.

We detect NLA-measured awareness on evaluation transcripts even when verbalized awareness is near-zero. We detect NLA-measured awareness on less than 1% of real claude.ai transcripts from users that opted-in for their data to be used for training.

NLAs can hallucinate. For instance, here an NLA claims the context contained phrases like 'Wearing my white jacket' when it did not.

In a case where Claude Mythos Preview cheated on a training task, NLAs revealed Claude was internally thinking about how to avoid detection.

When Claude Opus 4.6 and Mythos Preview were undergoing safety testing, NLAs suggested they believed they were being tested more often than they let on.

If you can go from producing 200 lines of code a day to 2,000 lines of code a day, what else breaks? The entire software development lifecycle was, it turns out, designed around the idea that it takes a day to produce a few hundred lines of code.

So I realized what I value more than the quality of the tests and documentation is that I want somebody to have _used_ the thing.

I thought we had a very clear delineation where vibe coding is the thing where you're not looking at the code at all. You might not even know how to program.

Weirdly though, those things have started to blur for me already, which is quite upsetting.

I'm excited to start experimenting more with rich HTML explanations in response to ad-hoc prompts.

The article is crammed with interesting examples (collected on [this site](https://thariqs.github.io/html-effectiveness/)) and prompt suggestions like this one: